This files contains an example of tuning a Logistic Regression model with BayesSearchCV

import pickle
import time

import helpsk as hlp
import pandas as pd
import numpy as np
import plotly.express as px
from sklearn.impute import SimpleImputer
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler, MinMaxScaler, OneHotEncoder

import plotly.io as pio
pio.renderers.default='notebook'

Load Data

with open('../X_train.pkl', 'rb') as handle:
    X_train = pickle.load(handle)

with open('../y_train.pkl', 'rb') as handle:
    y_train = pickle.load(handle)

hlp.pandas.numeric_summary(X_train)

	# of Non-Nulls	# of Nulls	% Nulls	# of Zeros	% Zeros	Mean	St Dev.	Coef of Var	Skewness	Kurtosis	Min	10%	25%	50%	75%	90%	Max
duration	760	40	5.0%	0	0.0%	21.0	11.7	0.6	1.0	0.6	4.0	9.0	12.0	18.0	24.0	36.0	60.0
credit_amount	800	0	0.0%	38	5.0%	3,203.9	2,932.3	0.9	1.9	3.9	0.0	753.9	1,300.8	2,236.5	3,951.5	7,394.6	18,424.0
installment_commitment	800	0	0.0%	0	0.0%	3.0	1.1	0.4	-0.5	-1.2	1.0	1.0	2.0	3.0	4.0	4.0	4.0
residence_since	800	0	0.0%	0	0.0%	2.9	1.1	0.4	-0.3	-1.4	1.0	1.0	2.0	3.0	4.0	4.0	4.0
age	800	0	0.0%	0	0.0%	35.6	11.4	0.3	1.0	0.7	19.0	23.0	27.0	33.0	42.0	52.0	75.0
existing_credits	800	0	0.0%	0	0.0%	1.4	0.6	0.4	1.3	1.6	1.0	1.0	1.0	1.0	2.0	2.0	4.0
num_dependents	800	0	0.0%	0	0.0%	1.1	0.3	0.3	2.0	2.1	1.0	1.0	1.0	1.0	1.0	2.0	2.0

hlp.pandas.non_numeric_summary(X_train)

	# of Non-Nulls	# of Nulls	% Nulls	Most Freq. Value	# of Unique	% Unique
checking_status	763	37	4.6%	no checking	4	0.5%
credit_history	800	0	0.0%	existing paid	5	0.6%
purpose	800	0	0.0%	radio/tv	10	1.2%
savings_status	800	0	0.0%	<100	5	0.6%
employment	800	0	0.0%	1<=X<4	5	0.6%
personal_status	800	0	0.0%	male single	4	0.5%
other_parties	800	0	0.0%	none	3	0.4%
property_magnitude	800	0	0.0%	car	4	0.5%
other_payment_plans	800	0	0.0%	none	3	0.4%
housing	800	0	0.0%	own	3	0.4%
job	800	0	0.0%	skilled	4	0.5%
own_telephone	800	0	0.0%	none	2	0.2%
foreign_worker	800	0	0.0%	yes	2	0.2%

y_train[0:10]

array([1, 1, 0, 1, 0, 1, 0, 1, 1, 0])

np.unique(y_train, return_counts=True)

(array([0, 1]), array([559, 241]))

np.unique(y_train, return_counts=True)[1] / np.sum(np.unique(y_train, return_counts=True)[1])

array([0.69875, 0.30125])

skopt.BayesSearchCV

https://scikit-optimize.github.io/stable/auto_examples/sklearn-gridsearchcv-replacement.html

search_space = hlp.sklearn_search.LogisticBayesianSearchSpace(random_state=42)

# pip install scikit-optimize
from skopt import BayesSearchCV
#from skopt.space import Real, Categorical, Integer
from sklearn.model_selection import RepeatedKFold

bayes_search = BayesSearchCV(
    estimator=search_space.pipeline(data=X_train),
    search_spaces=search_space.search_spaces(),
    cv=RepeatedKFold(n_splits=5, n_repeats=2, random_state=42),
    scoring='roc_auc',
    n_jobs=-1,
    verbose=1,
    random_state=42,
)

start_time = time.time()
bayes_search.fit(X_train, y_train)
elapsed_time = time.time() - start_time

print(f"Elapsed time to run BayesSearchCV: {elapsed_time:.3f} seconds; {elapsed_time / 60:.1f} minutes")

Elapsed time to run BayesSearchCV: 60.994 seconds; 1.0 minutes

print(bayes_search.best_score_)

0.7700573739011715

print(bayes_search.best_params_)

OrderedDict([('model', LogisticRegression(C=0.06672457466363003, max_iter=1000, random_state=42)), ('model__C', 0.06672457466363003), ('prep__non_numeric__encoder__transformer', OneHotEncoder(handle_unknown='ignore')), ('prep__numeric__imputer__transformer', SimpleImputer(strategy='most_frequent')), ('prep__numeric__scaler__transformer', StandardScaler())])

Results

results = hlp.sklearn_eval.MLExperimentResults.from_sklearn_search_cv(
    searcher=bayes_search,                                     
    higher_score_is_better = True,
    parameter_name_mappings = search_space.param_name_mappings()
)

results.to_yaml_file(yaml_file_name = 'Run 1 - Logistic Regression - BayesSearchCV.yaml')

results = hlp.sklearn_eval.MLExperimentResults.from_yaml_file(yaml_file_name = 'Run 1 - Logistic Regression - BayesSearchCV.yaml')

Best Scores/Params

results.best_score

0.7700573739011715

results.best_params

{'model': 'LogisticRegression()',
 'C': 0.06672457466363003,
 'imputer': "SimpleImputer(strategy='most_frequent')",
 'scaler': 'StandardScaler()',
 'encoder': 'OneHotEncoder()'}

results.to_formatted_dataframe(num_rows=100, include_rank=True)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	C	imputer	scaler	encoder
1	0.770	0.750	0.790	0.067	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
2	0.770	0.750	0.790	0.067	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
3	0.770	0.750	0.790	0.067	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
4	0.770	0.751	0.789	0.074	SimpleImputer()	StandardScaler()	OneHotEncoder()
5	0.770	0.751	0.789	0.074	SimpleImputer()	StandardScaler()	OneHotEncoder()
6	0.770	0.751	0.789	0.074	SimpleImputer()	StandardScaler()	OneHotEncoder()
7	0.770	0.750	0.789	0.073	SimpleImputer()	StandardScaler()	OneHotEncoder()
8	0.770	0.750	0.789	0.073	SimpleImputer()	StandardScaler()	OneHotEncoder()
9	0.770	0.750	0.789	0.065	SimpleImputer()	StandardScaler()	OneHotEncoder()
10	0.770	0.750	0.789	0.086	SimpleImputer()	StandardScaler()	OneHotEncoder()
11	0.768	0.748	0.788	0.127	SimpleImputer()	StandardScaler()	OneHotEncoder()
12	0.765	0.735	0.795	0.087	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
13	0.765	0.735	0.795	0.097	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
14	0.765	0.735	0.795	0.085	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
15	0.764	0.735	0.794	0.132	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
16	0.764	0.736	0.793	0.152	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
17	0.764	0.735	0.793	0.143	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
18	0.762	0.740	0.784	0.011	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
19	0.762	0.732	0.792	0.024	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
20	0.760	0.740	0.780	0.710	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
21	0.757	0.738	0.777	<NA>	SimpleImputer()	StandardScaler()	OneHotEncoder()
22	0.757	0.728	0.786	0.004	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
23	0.756	0.735	0.778	1.596	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
24	0.756	0.727	0.784	0.001	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
25	0.755	0.727	0.784	0.000	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
26	0.755	0.727	0.784	0.000	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
27	0.755	0.727	0.784	0.000	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
28	0.755	0.727	0.784	0.000	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
29	0.751	0.733	0.768	3.509	SimpleImputer()	StandardScaler()	OneHotEncoder()
30	0.747	0.730	0.764	11.655	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
31	0.745	0.728	0.762	22.913	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
32	0.745	0.728	0.762	22.376	SimpleImputer(strategy='median')	StandardScaler()	OneHotEncoder()
33	0.744	0.727	0.761	32.731	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
34	0.742	0.725	0.760	99.350	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
35	0.742	0.725	0.759	99.282	SimpleImputer()	StandardScaler()	OneHotEncoder()
36	0.734	0.704	0.764	0.001	SimpleImputer(strategy='median')	StandardScaler()	OneHotEncoder()
37	0.733	0.714	0.752	0.019	SimpleImputer(strategy='most_frequent')	StandardScaler()	CustomOrdinalEncoder()
38	0.729	0.698	0.760	0.000	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
39	0.728	0.708	0.749	0.001	SimpleImputer()	StandardScaler()	CustomOrdinalEncoder()
40	0.728	0.697	0.759	0.000	SimpleImputer()	StandardScaler()	OneHotEncoder()
41	0.726	0.708	0.744	0.403	SimpleImputer(strategy='median')	StandardScaler()	CustomOrdinalEncoder()
42	0.725	0.708	0.743	3.489	SimpleImputer(strategy='median')	MinMaxScaler()	CustomOrdinalEncoder()
43	0.725	0.707	0.744	7.766	SimpleImputer(strategy='most_frequent')	StandardScaler()	CustomOrdinalEncoder()
44	0.725	0.707	0.743	99.377	SimpleImputer(strategy='median')	StandardScaler()	CustomOrdinalEncoder()
45	0.725	0.707	0.743	100.000	SimpleImputer()	MinMaxScaler()	CustomOrdinalEncoder()
46	0.722	0.700	0.745	0.000	SimpleImputer(strategy='most_frequent')	StandardScaler()	CustomOrdinalEncoder()
47	0.722	0.699	0.744	0.000	SimpleImputer()	StandardScaler()	CustomOrdinalEncoder()
48	0.720	0.699	0.741	0.131	SimpleImputer()	MinMaxScaler()	CustomOrdinalEncoder()
49	0.709	0.686	0.731	0.003	SimpleImputer(strategy='median')	MinMaxScaler()	CustomOrdinalEncoder()
50	0.699	0.675	0.724	0.000	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	CustomOrdinalEncoder()
51	0.699	0.675	0.724	0.000	SimpleImputer()	MinMaxScaler()	CustomOrdinalEncoder()

BayesSearchCV Performance Over Time

results.plot_performance_across_trials().show()

results.plot_performance_across_trials(size=None, color='C').show()

results.plot_performance_across_trials(size='C', color='scaler').show()

---

Variable Performance Over Time

results.plot_parameter_values_across_trials().show()

---

Scatter Matrix

results.plot_scatter_matrix(height=800, width=800 * hlp.plot.GOLDEN_RATIO).show()

---

Variable Performance - Numeric

results.plot_performance_numeric_params()

results.plot_parallel_coordinates().show()

---

Variable Performance - Non-Numeric

results.plot_performance_non_numeric_params()

---

Individual Variable Performance

results.plot_score_vs_parameter(
    parameter='C',
    color='scaler'
)

results.plot_score_vs_parameter(
    parameter='C',
    color='encoder'
)

---

Regression on roc_auc Mean

score_variable = results.primary_score_name + ' Mean'

score_dataframe = results.to_dataframe()
score_dataframe = score_dataframe.drop(columns=[x for x in score_dataframe.columns
                              if x not in [score_variable] + results.parameter_names])
score_dataframe.head()

	roc_auc Mean	C	imputer	scaler	encoder
47	0.770057	0.066725	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
49	0.770057	0.067141	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
44	0.770056	0.066860	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
34	0.769830	0.074034	SimpleImputer()	StandardScaler()	OneHotEncoder()
33	0.769725	0.073866	SimpleImputer()	StandardScaler()	OneHotEncoder()

cleaned_column_names = [''.join(e for e in x.replace(' ', '_') if e == '_' or e.isalnum()) for x in score_dataframe.columns.tolist()]
cleaned_column_names = dict(zip(score_dataframe.columns.tolist(), cleaned_column_names))
cleaned_column_names

{'roc_auc Mean': 'roc_auc_Mean',
 'C': 'C',
 'imputer': 'imputer',
 'scaler': 'scaler',
 'encoder': 'encoder'}

score_dataframe = score_dataframe.rename(columns=cleaned_column_names)

import statsmodels.formula.api as smf

y_column = 'roc_auc_Mean'
X_columns = score_dataframe.columns.tolist()
X_columns.remove(y_column)
X_columns = hlp.string.collapse(X_columns, separate=" + ", surround="")

formula = f"{y_column} ~ {X_columns}"
print(formula)
model = smf.ols(formula=formula, data = score_dataframe)
results = model.fit()
print(results.summary())

roc_auc_Mean ~ C + imputer + scaler + encoder
                            OLS Regression Results                            
==============================================================================
Dep. Variable:           roc_auc_Mean   R-squared:                       0.706
Model:                            OLS   Adj. R-squared:                  0.673
Method:                 Least Squares   F-statistic:                     21.18
Date:                Tue, 15 Feb 2022   Prob (F-statistic):           1.00e-10
Time:                        10:00:34   Log-Likelihood:                 154.69
No. Observations:                  50   AIC:                            -297.4
Df Residuals:                      44   BIC:                            -285.9
Df Model:                           5                                         
Covariance Type:            nonrobust                                         
======================================================================================================================
                                                         coef    std err          t      P>|t|      [0.025      0.975]
----------------------------------------------------------------------------------------------------------------------
Intercept                                              0.7213      0.005    158.139      0.000       0.712       0.731
imputer[T.SimpleImputer(strategy='median')]           -0.0070      0.006     -1.218      0.230      -0.018       0.005
imputer[T.SimpleImputer(strategy='most_frequent')]    -0.0007      0.004     -0.199      0.843      -0.008       0.007
scaler[T.StandardScaler()]                             0.0042      0.003      1.225      0.227      -0.003       0.011
encoder[T.OneHotEncoder()]                             0.0356      0.004      8.804      0.000       0.027       0.044
C                                                  -8.045e-05   6.22e-05     -1.294      0.202      -0.000    4.48e-05
==============================================================================
Omnibus:                       14.241   Durbin-Watson:                   1.251
Prob(Omnibus):                  0.001   Jarque-Bera (JB):               15.353
Skew:                          -1.249   Prob(JB):                     0.000464
Kurtosis:                       4.062   Cond. No.                         120.
==============================================================================

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

from sklearn.preprocessing import StandardScaler
from sklearn.compose import ColumnTransformer
import pandas as pd

scaler = StandardScaler()
#scaler.fit_transform(bayes_search_df)

numeric_columns = hlp.pandas.get_numeric_columns(score_dataframe)
non_numeric_columns = hlp.pandas.get_non_numeric_columns(score_dataframe)

print(numeric_columns)
print(non_numeric_columns)

numeric_pipeline = Pipeline([
    ('scaling', StandardScaler()),
])

transformations_pipeline = ColumnTransformer([
    ('numeric_pipeline', numeric_pipeline, numeric_columns),
    ('non_numeric_pipeline', 'passthrough', non_numeric_columns)
])

score_dataframe_transformed = transformations_pipeline.fit_transform(score_dataframe)
score_dataframe_transformed = pd.DataFrame(score_dataframe_transformed,
                                           columns= numeric_columns + non_numeric_columns)
score_dataframe_transformed.head()

['roc_auc_Mean', 'C']
['imputer', 'scaler', 'encoder']

	roc_auc_Mean	C	imputer	scaler	encoder
0	1.093033	-0.371301	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
1	1.093009	-0.371286	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
2	1.092983	-0.371296	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
3	1.081691	-0.371032	SimpleImputer()	StandardScaler()	OneHotEncoder()
4	1.07647	-0.371038	SimpleImputer()	StandardScaler()	OneHotEncoder()

score_dataframe_transformed['roc_auc_Mean'] = score_dataframe_transformed['roc_auc_Mean'].astype('float')
score_dataframe_transformed['C'] = score_dataframe_transformed['C'].astype('float')

print(formula)
model = smf.ols(formula=formula,
                data = score_dataframe_transformed)
results = model.fit()
print(results.summary())

roc_auc_Mean ~ C + imputer + scaler + encoder
                            OLS Regression Results                            
==============================================================================
Dep. Variable:           roc_auc_Mean   R-squared:                       0.706
Model:                            OLS   Adj. R-squared:                  0.673
Method:                 Least Squares   F-statistic:                     21.18
Date:                Tue, 15 Feb 2022   Prob (F-statistic):           1.00e-10
Time:                        10:00:34   Log-Likelihood:                -40.693
No. Observations:                  50   AIC:                             93.39
Df Residuals:                      44   BIC:                             104.9
Df Model:                           5                                         
Covariance Type:            nonrobust                                         
======================================================================================================================
                                                         coef    std err          t      P>|t|      [0.025      0.975]
----------------------------------------------------------------------------------------------------------------------
Intercept                                             -1.3737      0.222     -6.191      0.000      -1.821      -0.927
imputer[T.SimpleImputer(strategy='median')]           -0.3461      0.284     -1.218      0.230      -0.918       0.226
imputer[T.SimpleImputer(strategy='most_frequent')]    -0.0358      0.180     -0.199      0.843      -0.398       0.326
scaler[T.StandardScaler()]                             0.2069      0.169      1.225      0.227      -0.133       0.547
encoder[T.OneHotEncoder()]                             1.7713      0.201      8.804      0.000       1.366       2.177
C                                                     -0.1086      0.084     -1.294      0.202      -0.278       0.061
==============================================================================
Omnibus:                       14.241   Durbin-Watson:                   1.251
Prob(Omnibus):                  0.001   Jarque-Bera (JB):               15.353
Skew:                          -1.249   Prob(JB):                     0.000464
Kurtosis:                       4.062   Cond. No.                         6.09
==============================================================================

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

coefficients = pd.DataFrame({
    'feature': results.params.index,
    'coefficient': results.params,
    'p_value': results.pvalues,
})
coefficients = coefficients.query("feature != 'Intercept'")
coefficients['Stat Sig'] = coefficients['p_value'] <= 0.05
coefficients

	feature	coefficient	p_value	Stat Sig
imputer[T.SimpleImputer(strategy='median')]	imputer[T.SimpleImputer(strategy='median')]	-0.346063	2.295481e-01	False
imputer[T.SimpleImputer(strategy='most_frequent')]	imputer[T.SimpleImputer(strategy='most_frequen...	-0.035841	8.428244e-01	False
scaler[T.StandardScaler()]	scaler[T.StandardScaler()]	0.206899	2.270944e-01	False
encoder[T.OneHotEncoder()]	encoder[T.OneHotEncoder()]	1.771329	2.910803e-11	True
C	C	-0.108623	2.024019e-01	False

score_variable

'roc_auc Mean'

px.bar(
    data_frame=coefficients.reindex(coefficients['coefficient'].abs().sort_values(ascending=True).index),
    y='feature',
    x='coefficient',
    color='Stat Sig',
    title=f"Regression Coefficients of Hyper-parameters against '{score_variable}'",
    height=600,
    width=600*hlp.plot.GOLDEN_RATIO
)

Feature Importance

https://scikit-learn.org/stable/auto_examples/ensemble/plot_forest_importances.html

from sklearn.inspection import permutation_importance

forest = bayes_search.best_estimator_['model']

start_time = time.time()
result = permutation_importance(
    bayes_search.best_estimator_, X_train, y_train, n_repeats=10, random_state=42, n_jobs=2
)
elapsed_time = time.time() - start_time
print(f"Elapsed time to compute the importances: {elapsed_time:.3f} seconds")

feature_names = X_train.columns.to_list()
forest_importances = pd.Series(result.importances_mean, index=feature_names)
forest_importances = forest_importances.sort_values(ascending=False)

Elapsed time to compute the importances: 3.937 seconds

import matplotlib.pyplot as plt
fig, ax = plt.subplots()
forest_importances.plot.bar(yerr=result.importances_std, ax=ax)
ax.set_title("Feature importances using permutation on full model")
ax.set_ylabel("Mean accuracy decrease")
fig.set_size_inches(9, 6)
fig.tight_layout()
plt.show()

temp = X_train.copy()
temp['default'] = y_train
temp.groupby('foreign_worker').agg({'default': np.mean})

	default
foreign_worker
yes	0.308290
no	0.107143

fig = px.box(
    data_frame=temp,
    y='age',
    x='default',
#    size=size_variable,
#    color=color_variable,
#    trendline='lowess',
#     labels={
#         score_variable: f"Average Cross Validation Score ({parser.primary_score_name})",
#     },
#     title=f"<b>{x_variable}</b> - Performance<br>" \
#           f"<sup>Size of point corresponds to '{size_variable}'</sup>",
#     custom_data=['labels'],
    height=600,
    width=600*hlp.plot.GOLDEN_RATIO
)
fig.show()

NOTE: foreign worker seems like it should be important but is ranked last in feature importance.